Microtexture Inpainting using Gaussian Models

Stochastic synthesis of micro-textures using ADSN model

Project Overview

This project focuses on micro-texture inpainting, where missing parts of images are reconstructed using stochastic modeling techniques. Micro-textures, which lack strong geometric patterns, can be effectively modeled as Gaussian random fields, enabling their synthesis through probabilistic methods.

Methodology

  • ADSN Model: Asymptotic Discrete Spot Noise for generating micro-textures
  • Kriging Conditioning: Ensures continuity at mask boundaries
  • Color Adaptation: Extension to color images while preserving inter-channel correlations
  • Covariance Matrix: Derived from auto-covariance function of micro-textures

Key Results

  • Grayscale and color micro-texture synthesis
  • Seamless inpainting with smooth boundary transitions
  • MSE at contour in order of 10^-21, demonstrating excellent continuity
  • Statistically consistent results

Technologies

  • Language: Python (Jupyter Notebook)
  • Libraries: NumPy, SciPy, Matplotlib
  • Methods: Gaussian random fields, Kriging interpolation, ADSN model

Project Report

View or download the full project report:

Microtexture Inpainting Report

Repository

View on GitHub